From: Masahiro TANAKA <masa16.tanaka@...>
Date: 2011-09-21T19:13:28+09:00
Subject: [ruby-core:39655] Re: [Ruby 1.9 - Bug #4576] Range#step miss the last value, if end-exclusive and has float number

2011/9/20 Tanaka Akira <akr@fsij.org>:
> However the algorithm doesn't solve [ruby-core:39602] and [ruby-core:39606].
>
> % ./ruby -e 'a = (1.0..12.7).step(1.3).to_a; p a.all? {|n| n <= 12.7 }, a.last'
> false
> 12.700000000000001

In my opinion, this behaviour is acceptable because the last value
generated by step method is close to the given Range-end value.

In the numerical calculation, the uniformity of sequence is more important.
Uniformity means that, if the generated sequence is
 { x(0), x(1), ..., x(n-1) },
then x(i+1)-x(i) is constant.

I think your solution [ruby-core:39612] is acceptable because the
modification of the last value is small.  If we need more uniformity of
the sequence, a possible algorithm is:

if (end < (n-1)*unit+beg) {
   for (i=0; i<n; i++) {
      rb_yield(DBL2NUM((n-1-i)/(n-1)*beg+i/(n-1)*end));
   }
} else ..


> % ./ruby -e 'e = 1+1E-12; a = (1.0 ... e).step(1E-16).to_a; p a.all?
> {|n| n < e }, a.last'
> false
> 1.000000000001

This problem is hard to solve because this is due to the accuracy of
floating point value.  The last part of this sequence is;

$  ruby -e 'e=1+1E-12; y=0; a=(1.0..e).step(1E-16).map{|x|"%.20f"%x};
p a[-6..-1]'
["1.00000000000099964481",
"1.00000000000099964481",
"1.00000000000099986686",
"1.00000000000099986686",
"1.00000000000100008890",
"1.00000000000100008890"]

The same value appears consecutively.  Therefore, even after the last
value is excluded, the last value is equal to the range-end.  This is
because this calculation exceeds the capability of floating point
arithmetic.  In my opinion, this case is not suitable for the test
case.  You can also see

$ ruby -e 'e=1+1E-12; y=0; a=(1.0..e).step(1E-16).map{|x|s=x-y;y=x;s};
p a[-6..-1]'
[2.220446049250313e-16, 0.0, 2.220446049250313e-16, 0.0,
2.220446049250313e-16, 0.0]

The difference is not equal to given step argument.  Even though
step*(n-1) == last-begin is still holds, This does not hold if you
decrease n, so I think the repeat times must not be decreased.

Masahiro Tanaka